Jan 10, 1992 - Journal of Pharmacokinetics and Biopharmaceutics, l/ol. 22, No. 4, 1994. Frequency Response Method in Pharmacokinetics. Ladislav Dedik ...
Journal of Pharmacokinetics and Biopharmaceutics, l/ol. 22, No. 4, 1994
Frequency Response Method in Pharmacokinetics Ladislav Dedik ~ and Mfiria I)uri~ovfi 2'3 Received January 10, 1992--Final June 6, .1994 The paper presents the demonstration of applicability of the fi'equency response method in a bioavailability study. The frequency response method, common in system engineering, is based on an approximation of the fi'equency response of a linear dynamic system, calculated from inputoutput measurements, by a fi'equency model of the system transfer function in the fi'equency domain, bz general, the influence of the system structure on the form of the system frequency response is much more distinct than on the form of the system output. Tbis is of great advantage in modeling the system fi'equency response instead of the system output, commonly used in pharmacokinetics. After a brief theoretical section, the method is demonstrated on the estimation of the rate and extent of gentamicin bioavailability after intratracheal administration to guinea pigs. The optimal fi'equency model of the system describing the gentamicin pathway into the systemic circulation and point estimates of its parameters were selected by the approximation of the system frequency response in the fi'equency domain, using a noniterative algorithm. Two similar estimates of the system weighting function were independently obtained: the weighting function of the selected fi'equency model and the weighting function estimated by the numerical deconvolution procedure. Neither of the estimates of the weighting function does decrease monotonously after the maximum of about 2.2-2.5 unit of dose. hr -~ recorded approximately O.1 hr after drug administration. Both esthnates show a marked additional peak approximately at 0.3 hr after administration and possible peaks in the further time period. We hypothesized that the loop found in the frequency response calculated and in the selected optimal fi'equency model, the high-order of this model, and several peaks identified in the estimates of the system weighting function indicated the complexity of the system and the presence of time delays. Three estimates of the extent of gentamicin intratracheal bioavailability obtained by the three different ways: directly from the calculated fi'equency response, calculated using the selected fi'equency model, and by the deconvolution method were 0.950, 0.934, and 0.907 respectively. Thus the conclusion can be made that gentamicin injected intratracheally to guinea pigs is almost completely available.
KEY WORDS: linear dynamic system; frequency response; frequency response method; weighting function; bioavailability; gentamicin.
This work was supported in part by Grant 283 from the Slovak Grant Agency, 814 38 Bratislava, Slovak Republic. tFaculty of Mechanical Engineering, Slovak Technical University, 812 31 Bratislava, Slovak Republic. 2Institute of Experimental Pharmacology, Slovak Academy of Sciences, 842 16 Bratislava, Slovak Republic. 3To whom correspondence should be addressed. 293 0090-466X/94/0800-0293507.00/0 9 1994Plenum Publishing Corporation
294
Dedik and l~urigov~
INTRODUCTION System modeling is usually done in the time domain, especially for pharmacokinetic purposes. Frequency domain methods are complementary to these procedures. The frequency response method is based on an approximation of a frequency response of a linear dynamic system measured or calculated from the system input-output measurements by a model of the system transfer function in the form of a ratio of polynomials in the complex domain (1). In spite of its positive features used to advantage, e.g., in system engineering or in some medical fields (2,3,4), the method is not commonly applied to pharmacokinetics. The frequency response method was introduced into pharmacokinetics several years ago (5,6,7). The utilization of the frequency response method is limited to the identification of pharmacokinetic models of linear systems consisting of subsystems without time delays, which are connected in serial (5). The frequency response method presented in our paper does not have such a limitation, and so it can be used for modeling any linear pharmacokinetic system consisting of several subsystems arranged in serial and/or parallel fashion, with or without time delays. The intent of our communication is to provide a methodological addition to previous papers (5,6,7) and to demonstrate the applicability of the frequency response method in a bioavailability study, using our own software CXT (Complex Tools for Linear Dynamic System Analysis), written in TURBO PASCAL for a personal computer. The example employed was that of gentamicin bioavailability assessment after intratracheal administration to guinea pigs. THEORETICAL Behaviour of a system with respect to a deterministic input signal is dependent on the static and dynamic properties of the system. If the harmonic signal Ci.(t)=Aj sin(D, t)
(1)
with the constant amplitude A~ and frequency D, is introduced into a linear dynamic system, the behaviour of the system in steady state is represented by the harmonic output signal
Cout(t) = Az(D ) sin(f~ - t - ~(f~ ))
(2)
with the amplitude A2(D ) and the lag ~(f~ ). The symbol t represents time. The complex function F(f~ ) = A2(f~ ) exp(-i 9 ~(f~ ))/d,
(3)
Frequency ResponseMethod
295
for the real argument f~e[O, ~ ) is the frequency response of the system. The symbol i represents the imaginary unit. If a deterministic nonperiodic signal is introduced into a linear dynamic system, the frequency response of the system can be determined as F(12 )
=
Co,t(f~ )/C~,(f~ )
(4)
for the real argument l'~a[D-mi,,,f~max]- C~,(f~) and Co,t(~) are the Fourier transforms of the system input and output signal, respectively. The system frequency response in the frequency domain, or the system weighting function in the time domain (8,9) provides all characteristics of the linear dynamic system. In the frequency domain, the system static properties are represented by the F(0) value and dynamic properties by the dimensionless function F(f~ ) F(E~ ) = F(Ft )/F(0), (F(0) = 1)
(5)
called the normalized frequency response of the linear dyanmic system. In the time domain, the static and dynamic properties of the system are represented by the area under the system weighting function and by the weighting function, respectively. One of the conventional plots of the system frequency response is its polar plot in the complex plane. As an example the inherent frequency response of the first-order linear dynamic system described by Eqs. (6) and
(7) G H(s) = - -
l+T's
Co,t(t) + T . dC.ut(t) _ G" Q , ( t )
(6) (7)
dt
in the Laplace s domain and in the time domain, respectively, is presented in Fig. 1. G is the system gain and T is the system time constant. The form of the system frequency response is strongly dependent on the system structure. In general, the influence of the system structure on the form of the system frequency response is much more distinct than on the form of the system output. This is of great advantage in modeling the system frequency response instead of the system output, commonly used in pharmacokinetics (10). A linear pharmacokinetic system whose frequency response has been calculated from the input-output measurements according to Eq. (4), can be described by the ferquency model whose properties are similar to those of the calculated frequency response. The model transfer function HM(S)
296
Dedik and l~uri~ov~
COMPLEX PLANE IMAGINARY AXIS
+i
(0,0)
(G,O) REAL AXIS
J
-i Fig. 1. Polar plot of the frequencyresponse of the first-order linear dynamic system,defined by Eq. (6) and/or (7), in the complexplane. may have the form
ttM(S) =/qM(S) " G
(8)
where HM(S) is the model of the normalized frequency response, i.e., the model of the system dynamic properties and the gain G is the estimator of the static model parameter. The linear pharmaeokinetic system that does not contain a time delay can be approximated by the model
~I~(s) = A(s)/~(s)
(9)
where A(s) and B(s) are polynomials (I). The linear pharmacokinetic system that contains a time delay can be approximated by the same model, however, only by high-order numerator and demoninator polynomials with complex conjugated roots (11). The system frequency model HM(~ ) can be formally obtained by the substitution s = i. ~ into Eqs. (8) and (9). The optimal frequency model, optimal frequency range, G value, and values of the polynomial coefficients
FrequencyResponseMethod
297
can be selected by fitting variable frequency models HM(~'~) to the calculated frequency response of the pharmacokinetic system (1). Since F(f~ ) is a complex function of the real argument f~, the imaginary unit i is not stated in its symbolic representation. MATERIAL AND METHODS The published plasma concentrations of gentamicin were used (10). Gentamicin was administered intravenously and intratracheally in the same doses to guinea pigs (10). The drug plasma concentrations were determined in samples taken from the left side of the heart. System Equations The plasma circulatory system after intravenous administration was described by Civ(S) = H e ( s ) ' / i v ( S )
(10)
where/iv(s) was the intravenous input, ltc(s) and Civ(s) were the corresponding transfer function and system output, respectively. The plasma circulatory system after intratracheal administration was described by Cit(s) = He(s) 9 Hit(s)"/it(s)
(11)
where/it(s) was the intratracheal input, the product of H~(s) 9 Hit(s) and Cit(s) were the corresponding transfer function and system output, respectively. For the same inputs,/iv(S) =/it(s) (10), Hit(s) was expressed as Hit(s) = fit (s)//fly(S)
(12)
In this case, Cit(s) and Civ(S) were the respective output and input functions of the system defined by Eq. (12). Hit(s) was the system transfer function and the corresponding inverse transform Hit(t) was the weighting function of the system (1). The arguments presented were based on the assumption that the disposition kinetics of gentamicin in guinea pigs was linear. Linearity was assumed in the sense that all responses adhered to the superposition principle and time invariance with respect to the input rate. Frequency Response Method The frequency responses Fit(~r~) of the system defined by Eq. (12) was calculated according to Eq. (4) for varying sets of the fl values. Each set
298
Dedik and Duri~ov~
contained N = 20 values of f2, ranging from ~')-minto f2. . . . generated by the geometric series. The Cit(f~ ) and Civ(f~) functions were determined as the Fourier transforms of the straight lines approximating the corresponding drug concentrations between two adjacent sampling points and as the Fourier transforms of the exponentials approximating the concentrationtime courses in the interval from the last sampling time to infinity. F~t(0) was determined as G ( 0 ) = lim F~t(n )
(13)
~0
where lim F,(f~ ) = A UCit/A UCiv
fl---,0
with A UCit and A UCiv representing the areas under the plasma concentration-time curves from time 0 to infinity after intratracheal and intravenous administration, respectively9 The Fit(0~ value was used for the calculation of the normalized frequency response Fit(f~ ) according to Eq. (5)9 To approximate the normalized frequency response F~t(~ ) of the system defined by Eq. (12) the following frequency model was employed
/~M,it(~.-~j)
-
-
ao + at i~j + a2( i~j) 2 +" 9 9 + a,( i~i)" l + b t i f ~ i + b 2 ( i ~ i ) 2 + ' " + b m ( i f 2 j ) ''
(14)
The point estimates of the polynomial coefficients of variable frequency models with varying polynomial degrees n and m were obtained by fitting the respective models/tM,it(f2 ) to the calculated ~t(fl ), using the noniterative algorithm in the complex domain, based on the Levy method (12). The formal description of this algorithm, used in the CXT program was as follows (13): J~ ((~1, i~']j) = ao + al i n i + a2( ifli) 2 +" 9 9 + a,,( ifli)"
f2( 02, i~j) = b, i~j + bz( i~j) 2 + . . 9+ b,n( i~j)"
where ao
al
an
(i b2
n
Frequency Response Method
299
were the vectors of unknown parameters. The j value of the calculated normalized frequency response ~ , j ~ [ l , N] was expressed as
fi,/_ f l ( O l , iD/)
i j)
(15)
where Zj was the random variable describing the error. Eq. (15) was rewritten into Eq. (16)
~/=f~ (|
i~2j) - Fi 9f2(|
in/) q- Zj
(16)
under the condition that the term Z/" J~(O2, iD/) converged to zero (12,13). The minimal frequency Dmi, was selected according to the empirical condition lag ~(Dmi,) = 7r/120 rad (13). To decide on the optimal Dmaxvalue and on the numerator and denominator polynomial degrees of the frequency model optimally describing/~it(D) the criterion
CC = 2. N ln(Rrea~ + Rimag) "+-2" (n + m + 1)
(17)
was applied, assuming the normal distribution of the random variable Zy. Rreat and R~,l,a~ represented the real and imaginary parts of the residual sum of squares, respectively. To test the null hypothesis that the observed distribution of the residuals Zj,re~1, Zj,im~g can be regarded as the normal distribution, the Z 2 test was used (14). The frequency model with minimum CC was regarded as the best representation of the system frequency response. The static model parameter Git w a s estimated as Git = a0" El,(0)
(18)
The weighting function HM,it(t) of the selected optimal frequency model was obtained as the response of this model to the Dirac-delta pulse using the numerical Euler method (13). Numerical Deconvolution Method
The numerical deconvolution method (15) was modified as follows: The initial drug concetrations in the plasma circulatory system were assumed to be C,v(t = o) = ci,(t = o) = o
(19)
The interpolated points obtained by Lagrange cubic polynomials fitted to the logarithms of the measured concentration data were used. The step of the interpolation was equal to the first sampling time. The static parameter of the system defined by Eq. (12) was estimated as the area under weighting function in the time interval from the first to last sampling time, using the
300
Dedik and l~uri~v~
linear trapezoidal method. The first and last sampling time intervals were the same in both modes of administration (10). The calculations were done by the TURBO PASCAL programs: CXT and DND (Direct Numerical Deconvolution); BIO-LAB Bratislava. (DEMO versions of the programs will be sent on receipt of a floppy in a self-addressed floppy disk mailer.) RESULTS The points in Fig. 2 show the calculated normalized frequency response of the system defined by Eq. (12) determined for the selected interval from f~m~,=0.058 r a d - h r -I to flmax=30.42 rad" hr-L The selected optimal frequency model had the numerator and denominator polynomial degrees n = 4 and m = 5, respectively. It is presented as the solid line in Fig. 2. The calculations of the goodness-of-fit statistics yielded the dispersion values 2 orc,~ and O'~magof 0.00037 and 0.00016, respectively. The g 2 test of the null IMAGINARY AXIS
0.25i
-0.5
(o,o)
I
/
I
REAL A
t
-0.75i Fig. 2. Calculated normalized frequency response (points) of the system defined by Eq. (12) for the selected interval o f f l values from f~mi,=0.058 rad.hr ' to f~max=30.42 rad.hr -4. Model approximation by the selected optimal frequency model (solid line) with the numerator and denominator polynomial degrees n = 4 and m = 5, respectively.
iJ
Frequency Response Method
301
Table I. Point Estimates of the Polynomial Coefficientsof the Optimal
Frequency Model of the System Defined by Eq. (12) Numerator polynomial Denominatorpolynomial A(s) B(s) Dimension ao= 0.983 al =0.127 a2= 0.014 a3 = 0.000597 a4 = 0.0000176 --
-bl =0.477 b2= 0.0661 b3= 0.0087 b4= 0.000275 b5= 0.0000177
-hr hr 2
hr3 hr4 hr5
hypothesis that the observed distribution of the corresponding residuals Zj.re~l, Zj.imagcan be regarded as the normal distribution yielded the following values: Z2(3)= 7.0232, P=0.0711. Since the probability level was greater than the conventional 0.05 level, the null hypothesis that the distribution of the observed residuals was normal was not rejected. The point estimates of the fitted coefficients of the numerator and denominator polynomials of the selected optimal model are assembled in Table I. The calculation o f the roots of the numerator (rj, r2, r3, r 4 ) and denominator (qJ,q2,q3,q4,qs) polynomials yielded the values of 1"1,2= - 12.319 4- 16.938i, r3,4 = -4.654 4- 10.283i, and those of ql,2 = -3.477 4- 18.635i, q3.4= -2.866 4- 6.892i, qs = -2.818, respectively. The loop of the frequency response, determined for the high values of ~ , seen in the bottom left quadrant of the complex plane in Fig. 2, the high order of the selected optimal frequency model, and presence of the complex conjugated roots indicate the complexity of the system and the presence of system time delays (1,11 ). The open squares in Fig. 3a illustrate the mean concentrations of gentamicin in plasma after its intravenous administration, representating the input of the system defined by Eq. (12). The filled squares in Fig. 3b show the time course of the mean concentrations of gentamicin in plasma after its intratracheal administration, representing the output of the system defined by Eq. (12). The approximations of this time course by the output of the selected optimal frequency model and by the three-exponential function published in our previous study (10) are shown as the solid and dashed line, respectively, in Fig. 3b. The estimate of the weighting function of the selected optimal frequency model and the estimate obtained by the deconvolution method are shown as the solid and dashed line, respectively, in Fig. 4. The estimates, approaching the rate of gentamicin bioavailability after intratracehal administration, did not decrease monotonously after their maxima, of about 2.2-2.5 unit of dose. hr -~, recorded approximately 0.1 hr after administration, but showed marked additional peaks approximately at 0.3 hr after administration, and
302
Dedik and l~uri~ov~
30-
Z
O a
a::' 20P" E
13
Q
LIJ
O --- lOz O o TIME ( h r ) 15-
•
